Log CUDA kernel API calls and inputs to diagnose SGLang crashes.
npx skills add https://github.com/sqjian/sglang --skill debug-cuda-crash-sqjian
Please help me install this Agent Skill. Skill: debug-cuda-crash Source: https://github.com/sqjian/sglang/tree/main/.claude/skills/debug-cuda-crash Command: npx skills add https://github.com/sqjian/sglang --skill debug-cuda-crash-sqjian
This skill enables debugging CUDA crashes in SGLang by capturing kernel API call data and input tensors before a crash to preserve debugging context and reproduce failures.
Enable kernel API logging at Level 3 or higher and run your script to reproduce and inspect CUDA crashes.
High-intent search queries and answers about installing and using this skill.
To diagnose a CUDA crash in SGLang, enable kernel API logging at Level 3 or higher to record function boundaries and inputs. This captures invalid inputs or misaligned tensors before the failure.
Kernel API logging records function boundaries, input tensors, and optional metadata to preserve debugging context. It identifies invalid inputs or misaligned tensors causing CUDA device-side assertions during workloads.
Optional API filtering allows you to isolate specific kernel calls during CUDA crash diagnostics. Combined with configurable log destinations and dump directories, this enables reproducible debugging across processes.
Kernel API logging handles both LLM and diffusion workloads in SGLang. It captures input and output metadata to locate invalid inputs causing CUDA device-side assertions across these different model types.
Crash-safe dumps preserve debugging context by writing kernel API inputs and metadata to configured dump directories before the crash. This ensures multi-process logs survive for inspecting invalid tensors.
annealing-inversionEnable Kernel API logging to diagnose CUDA crashes in SGLang.
QingHuan-6Diagnose CUDA crashes in SGLang by exposing kernel API call traces and tensor metadata.
FutureMLS-LabDiagnose SGLang CUDA crashes by recording kernel API calls and tensor metadata.
spacecat2002Diagnose CUDA crashes in SGLang by logging tensors at kernel API boundaries.
